[bittorrent] State of Super-seeding

Arnaud Legout arnaud.legout at sophia.inria.fr
Thu Aug 9 11:40:31 EDT 2007


tiennou wrote:
>   I'm working on implementing Superseeding and the Fast Peers 
> Extension in Transmission, and I find the spec lacking some information :
> - Is it really useful to implement Superseeding ? (There's a post in 
> the archives saying that the updated choke algorithm was now sufficient.)

I don't see any strong evidence that superseeding makes sense.
You can surely construct pathological cases in which your initial seed has
a very low upload capacity to justify superseeding. However, I have never
seen any rigorous measurement that shows a large overhead with either
version of choking algorithm in seed state.

Our measurements have never shown an overhead much larger than 10%.

As long as rarest first is correctly implemented, I see no reason to 
implement superseeding.
By correctly implemented I mean in particular that when you have more 
that one piece in
your rarest pieces set, you need to select the next piece to request at 
random in this rarest set
The common mistake is to select the piece in a deterministic order (for 
instance based on the sequence number)
that is the same among all peers. The result is that when a torrent is 
starting, all peers see the same rarest pieces, therefore
they all download the same piece as long as one peer got this piece 
(which make the piece no longer a rarest one).
Such an implementation mistakes may lead to a large overhead at the 
initial seed in the torrent start phase.


More information about the BitTorrent mailing list